HUE Vision

Specification status: SPEC Product status: TBI Open implementation choices: TBD

North star

HUE is an open-source, local-first personal agent workspace: an AI work operating system for projects and ongoing areas of life. It combines the conversational ease of a general assistant, the organization of Projects and permanent Areas, the portability of a second brain, the repository fluency of coding agents, and the visible delegation of a capable team—without making the user administer models, agents, providers, terminals and chat histories manually.

The central idea is not one all-knowing agent. The workspace establishes the correct context, memory, sources, permissions, tools and specialist for each independent session. Its durable “brain” is structured user-owned knowledge, current space state, decisions, source bindings and retrieval rules—not the context window of Hermes, OpenCode or any particular model.

The desired experience is simple:

Open the right Project or Area, begin or resume a session, and let HUE establish the context and execution environment automatically; watch work progress, intervene when useful, and review verified results in context.

The missing product

Today, users assemble fragments:

• A chat application understands general requests but has shallow project folders and little execution visibility.
• A coding agent works well inside one repository but is not a complete workspace for life, research, documents and native applications.
• An agent framework can coordinate workers but expects its user to be a software architect.
• A personal assistant has memory and tools but often hides orchestration, mixes contexts, or requires manual profile selection.
• A computer-use product can click through applications but does not own project history, code, tasks, memory and artifacts.

HUE makes those fragments one coherent product.

Final product promise — TBI

When implemented, HUE will provide:

1. One workspace, many independent contexts. HUE provides one coherent front door while each session receives only the relevant global preferences and Space context.
2. Projects, Areas and Resources. Projects pursue finishable outcomes; Areas represent ongoing responsibilities; Resources provide reusable reference material with explicit relationships and access.
3. Real spaces. Each Project or Area binds goals, a human-readable context pack, source systems, files, knowledge, decisions, permissions, memories, sessions, tasks, artifacts and defaults.
4. Sessions and execution together. Discussion, execution, research, monitoring and review sessions stay independent; any session can create or connect to a durable task without switching products.
5. Visible orchestration. A bounded scheduler/router decomposes work, creates temporary specialists, shows dependencies and progress, and revises the plan when evidence changes.
6. Autonomous resource selection. HUE chooses an appropriate model, provider, effort, toolset, runtime and isolation level through transparent policy.
7. OpenCode-first, replaceable development. OpenCode is the initial primary software execution environment; worktrees, edits, tests, diffs, reviews and PR handoff remain defined by HUE contracts so other runtimes can replace or complement it.
8. Cowork-class knowledge work. Files, documents, spreadsheets, web research, browser activity and native computer use belong to the same workspace and task system.
9. A portable knowledge substrate. Human-readable context packs, explicit decisions, backlinks, tags, provenance and maintained summaries remain useful without the AI layer.
10. Layered memory. Session, Space, global preference, source and episodic archive layers remain distinct, inspectable and correctable.
11. Source ownership. GitHub, Calendar, email and user files remain authoritative for their native objects; HUE synchronizes projections and provides a control surface rather than creating stale parallel truth.
12. Durable work. Sessions and tasks survive page refreshes and process restarts; users can pause, resume, steer, branch and cancel.
13. Proof, not claims. Results link to changed files, test output, citations, recordings, screenshots and decision logs.
14. Human authority. Sensitive, external, irreversible and costly operations are policy-controlled and approval-gated.
15. Open and adaptable. HUE can use different model providers, worker runtimes, tool protocols and local or hosted infrastructure. Hermes and OpenCode sit behind HUE’s abstraction.

Mental model

HUE                         the personal workspace and control surface
Space                       a Project or Area context + permission boundary
Project                     a Space pursuing a finishable outcome
Area                        a Space representing an ongoing responsibility
Resource                    reusable reference linked to Spaces explicitly
Session                     an independent discussion/execution/research/monitor/review context
Conversation                the message stream inside a Session
Task                        a durable desired outcome with lifecycle
Run                         one execution attempt for a task
Plan                        an evolving graph of steps and dependencies
Worker                      a temporary specialist execution context
Tool                        a bounded capability invoked by a worker
Approval                    a human decision at a policy boundary
Artifact                    a durable output or evidence object
Knowledge/context pack      portable maintained state for a Space
Memory                      a curated fact/decision available to later work
Event                       an immutable record of what occurred during a run

The intended relationship

flowchart TD
    U[User] --> H[HUE personal workspace]
    H --> G[Global preferences]
    H --> P[Selected Space: Project or Area]
    P --> C[Independent Sessions]
    P --> T[Durable tasks]
    P --> F[Files, folders and repositories]
    P --> M[Context pack, knowledge and decisions]
    P --> R[Permissions and resource policy]
    T --> O[Orchestrator]
    O --> W1[Temporary researcher]
    O --> W2[Temporary developer]
    O --> W3[Temporary reviewer]
    O --> W4[Browser or computer operator]
    W1 --> E[Evidence and artifacts]
    W2 --> E
    W3 --> E
    W4 --> E
    E --> H
    H --> U

Product posture

HUE should feel like a calm studio and personal operating workspace, not an agent command center by default.

• The user sees outcomes, blockers, decisions and meaningful progress.
• The system keeps raw event detail available without forcing it into the primary conversation.
• Advanced users can inspect or override routing, models, tools and execution topology.
• New users can simply ask for work and understand what HUE is doing.

What HUE is not

• Not a generic multi-agent SDK with a chat page attached.
• Not one omniscient assistant context expected to know every codebase and aspect of life.
• Not a thin wrapper around one model provider or agent framework.
• Not a collection of separately operated specialist bots.
• Not a project-management system that requires users to hand-maintain every task.
• Not an IDE replacement; it cooperates with editors and repositories while owning AI work context and execution.
• Not an unrestricted autonomous employee. Human policy, permissions and review remain fundamental.
• Not a system that calls an LLM for deterministic work a normal program can do better.

Success criteria

HUE succeeds when a user can:

• move between unrelated Projects and Areas without context leakage or manual reconfiguration;
• run independent sessions concurrently without collapsing them into one context window;
• assign a complex outcome without selecting a model, provider, effort, agent type or directory;
• understand the plan and current status in under ten seconds;
• safely leave work running and later resume from a truthful durable record;
• review exactly what changed and why;
• correct memory and routing behavior without editing hidden prompts;
• use local data and local models where required while retaining optional cloud capability;
• replace several disconnected AI applications with one coherent workspace.

Final review question

Every proposed feature should answer:

Does this help HUE place an independent session inside the correct Space, assemble the right user-owned context, operate safely, organize work visibly, and return evidence-backed outcomes with less manual agent administration?

If not, it probably does not belong in the core product.